File size: 11,047 Bytes
2bbfbb7 0393dfa 2bbfbb7 668baf3 2bbfbb7 0393dfa 2bbfbb7 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 |
//! Speaker and emotion embedding models
use crate::{Error, Result};
use ndarray::{Array1, Array2, Array, IxDyn};
use std::collections::HashMap;
use std::path::Path;
use super::OnnxSession;
/// Speaker encoder for extracting speaker embeddings from audio
pub struct SpeakerEncoder {
session: Option<OnnxSession>,
embedding_dim: usize,
}
impl SpeakerEncoder {
/// Load speaker encoder from ONNX model
pub fn load<P: AsRef<Path>>(path: P) -> Result<Self> {
let session = OnnxSession::load(path)?;
Ok(Self {
session: Some(session),
embedding_dim: 192, // CAMPPlus default
})
}
/// Create placeholder encoder (for testing)
pub fn new_placeholder(embedding_dim: usize) -> Self {
Self {
session: None,
embedding_dim,
}
}
/// Extract speaker embedding from mel spectrogram
pub fn encode(&self, mel_spectrogram: &Array2<f32>) -> Result<Array1<f32>> {
if let Some(ref session) = self.session {
// Prepare input (add batch dimension)
let input = mel_spectrogram
.clone()
.into_shape(IxDyn(&[1, mel_spectrogram.nrows(), mel_spectrogram.ncols()]))?;
let mut inputs = HashMap::new();
inputs.insert("mel".to_string(), input);
let outputs = session.run(inputs)?;
let embedding = outputs
.get("embedding")
.ok_or_else(|| Error::Model("Missing embedding output".into()))?;
// Extract 1D embedding
let flat: Vec<f32> = embedding.iter().cloned().collect();
Ok(Array1::from_vec(flat))
} else {
// Return random embedding for testing
Ok(Array1::from_vec(vec![0.0f32; self.embedding_dim]))
}
}
/// Extract embedding from audio file
pub fn encode_audio(&self, audio_path: &str) -> Result<Array1<f32>> {
use crate::audio::{compute_mel_from_file, AudioConfig};
let config = AudioConfig::default();
let mel = compute_mel_from_file(audio_path, &config)?;
self.encode(&mel)
}
/// Get embedding dimension
pub fn embedding_dim(&self) -> usize {
self.embedding_dim
}
/// Normalize embedding to unit length
pub fn normalize_embedding(&self, embedding: &Array1<f32>) -> Array1<f32> {
let norm = embedding.iter().map(|x| x * x).sum::<f32>().sqrt();
if norm > 1e-8 {
embedding / norm
} else {
embedding.clone()
}
}
/// Compute cosine similarity between embeddings
pub fn cosine_similarity(&self, emb1: &Array1<f32>, emb2: &Array1<f32>) -> f32 {
let norm1 = emb1.iter().map(|x| x * x).sum::<f32>().sqrt();
let norm2 = emb2.iter().map(|x| x * x).sum::<f32>().sqrt();
if norm1 < 1e-8 || norm2 < 1e-8 {
return 0.0;
}
let dot: f32 = emb1.iter().zip(emb2.iter()).map(|(a, b)| a * b).sum();
dot / (norm1 * norm2)
}
}
/// Emotion encoder for controlling emotional expression
pub struct EmotionEncoder {
/// Emotion embedding matrix (num_emotions x embedding_dim)
emotion_matrix: Array2<f32>,
/// Number of emotion dimensions
num_dims: usize,
/// Values per dimension
dim_sizes: Vec<usize>,
}
impl EmotionEncoder {
/// Create emotion encoder with specified dimensions
pub fn new(num_dims: usize, dim_sizes: Vec<usize>, embedding_dim: usize) -> Self {
let total_emotions: usize = dim_sizes.iter().sum();
let emotion_matrix = Array2::zeros((total_emotions, embedding_dim));
Self {
emotion_matrix,
num_dims,
dim_sizes,
}
}
/// Load emotion matrix from file
pub fn load<P: AsRef<Path>>(path: P) -> Result<Self> {
let path = path.as_ref();
if !path.exists() {
return Err(Error::FileNotFound(path.display().to_string()));
}
// Load safetensors file
let file_data = std::fs::read(path)?;
let tensors = safetensors::SafeTensors::deserialize(&file_data)
.map_err(|e| Error::ModelLoading(format!("Failed to load safetensors: {}", e)))?;
// Extract emotion matrix
let tensor = tensors
.tensor("emotion_matrix")
.map_err(|e| Error::ModelLoading(format!("Missing emotion_matrix: {}", e)))?;
let shape = tensor.shape();
let data: Vec<f32> = tensor.data().chunks_exact(4).map(|b| {
f32::from_le_bytes([b[0], b[1], b[2], b[3]])
}).collect();
if !tensor.data().chunks_exact(4).remainder().is_empty() {
return Err(Error::ModelLoading("Tensor data length is not a multiple of 4".to_string()));
}
let emotion_matrix = Array2::from_shape_vec((shape[0], shape[1]), data)
.map_err(|e| Error::ModelLoading(format!("Shape mismatch: {}", e)))?;
// Default configuration
let num_dims = 8;
let dim_sizes = vec![5, 6, 8, 6, 5, 4, 7, 6];
Ok(Self {
emotion_matrix,
num_dims,
dim_sizes,
})
}
/// Encode emotion vector to embedding
pub fn encode(&self, emotion_vector: &[f32]) -> Result<Array1<f32>> {
if emotion_vector.len() != self.num_dims {
return Err(Error::ShapeMismatch {
expected: format!("{} dimensions", self.num_dims),
actual: format!("{} dimensions", emotion_vector.len()),
});
}
let embedding_dim = self.emotion_matrix.ncols();
let mut embedding = vec![0.0f32; embedding_dim];
let mut offset = 0;
for (WIN_LENGTH, (&value, &dim_size)) in emotion_vector.iter().zip(self.dim_sizes.iter()).enumerate() {
// Interpolate between discrete emotion levels
let continuous_idx = value * (dim_size - 1) as f32;
let lower_idx = continuous_idx.floor() as usize;
let upper_idx = (lower_idx + 1).min(dim_size - 1);
let alpha = continuous_idx - lower_idx as f32;
// Weighted combination
for i in 0..embedding_dim {
let lower_val = self.emotion_matrix[[offset + lower_idx, i]];
let upper_val = self.emotion_matrix[[offset + upper_idx, i]];
embedding[i] += lower_val * (1.0 - alpha) + upper_val * alpha;
}
offset += dim_size;
}
// Normalize
let norm: f32 = embedding.iter().map(|x| x * x).sum::<f32>().sqrt();
if norm > 1e-8 {
for e in embedding.iter_mut() {
*e /= norm;
}
}
Ok(Array1::from_vec(embedding))
}
/// Get neutral emotion (all zeros)
pub fn neutral(&self) -> Vec<f32> {
vec![0.5f32; self.num_dims]
}
/// Get preset emotion vectors
pub fn preset(&self, name: &str) -> Vec<f32> {
match name {
"happy" => vec![0.9, 0.7, 0.6, 0.5, 0.5, 0.5, 0.5, 0.5],
"sad" => vec![0.2, 0.3, 0.4, 0.5, 0.6, 0.5, 0.5, 0.5],
"angry" => vec![0.8, 0.9, 0.7, 0.5, 0.3, 0.5, 0.5, 0.5],
"fearful" => vec![0.3, 0.4, 0.8, 0.5, 0.7, 0.5, 0.5, 0.5],
"surprised" => vec![0.7, 0.8, 0.7, 0.5, 0.5, 0.5, 0.5, 0.5],
"neutral" | _ => self.neutral(),
}
}
/// Interpolate between two emotion vectors
pub fn interpolate(&self, emot1: &[f32], emot2: &[f32], alpha: f32) -> Vec<f32> {
emot1
.iter()
.zip(emot2.iter())
.map(|(&a, &b)| a * (1.0 - alpha) + b * alpha)
.collect()
}
/// Apply emotion strength/alpha
pub fn apply_strength(&self, emotion: &[f32], strength: f32) -> Vec<f32> {
let neutral = self.neutral();
self.interpolate(&neutral, emotion, strength)
}
}
/// Semantic encoder for extracting semantic codes
pub struct SemanticEncoder {
session: Option<OnnxSession>,
embedding_dim: usize,
}
impl SemanticEncoder {
/// Load semantic encoder
pub fn load<P: AsRef<Path>>(path: P) -> Result<Self> {
let session = OnnxSession::load(path)?;
Ok(Self {
session: Some(session),
embedding_dim: 1024,
})
}
/// Create placeholder encoder
pub fn new_placeholder() -> Self {
Self {
session: None,
embedding_dim: 1024,
}
}
/// Encode audio to semantic codes
pub fn encode(&self, audio: &[f32], sample_rate: u32) -> Result<Vec<i64>> {
if let Some(ref session) = self.session {
let input = Array::from_shape_vec(
IxDyn(&[1, audio.len()]),
audio.to_vec(),
)?;
let mut inputs = HashMap::new();
inputs.insert("audio".to_string(), input);
let outputs = session.run(inputs)?;
let codes = outputs
.get("codes")
.ok_or_else(|| Error::Model("Missing codes output".into()))?;
Ok(codes.iter().map(|&x| x as i64).collect())
} else {
// Return dummy codes for testing
let num_codes = audio.len() / (sample_rate as usize / 50); // ~50 codes/sec
Ok(vec![0i64; num_codes.max(1)])
}
}
}
#[cfg(test)]
mod tests {
use super::*;
#[test]
fn test_speaker_encoder_placeholder() {
let encoder = SpeakerEncoder::new_placeholder(192);
assert_eq!(encoder.embedding_dim(), 192);
}
#[test]
fn test_emotion_encoder() {
let encoder = EmotionEncoder::new(8, vec![5, 6, 8, 6, 5, 4, 7, 6], 256);
let neutral = encoder.neutral();
assert_eq!(neutral.len(), 8);
assert!(neutral.iter().all(|&x| (x - 0.5).abs() < 1e-6));
}
#[test]
fn test_emotion_presets() {
let encoder = EmotionEncoder::new(8, vec![5, 6, 8, 6, 5, 4, 7, 6], 256);
let happy = encoder.preset("happy");
assert_eq!(happy.len(), 8);
assert!(happy[0] > 0.5); // Happy has high first dimension
}
#[test]
fn test_emotion_interpolation() {
let encoder = EmotionEncoder::new(8, vec![5, 6, 8, 6, 5, 4, 7, 6], 256);
let happy = encoder.preset("happy");
let sad = encoder.preset("sad");
let mid = encoder.interpolate(&happy, &sad, 0.5);
// Middle value should be average
for i in 0..8 {
assert!((mid[i] - (happy[i] + sad[i]) / 2.0).abs() < 1e-6);
}
}
#[test]
fn test_cosine_similarity() {
let encoder = SpeakerEncoder::new_placeholder(3);
let emb1 = Array1::from_vec(vec![1.0, 0.0, 0.0]);
let emb2 = Array1::from_vec(vec![1.0, 0.0, 0.0]);
let sim = encoder.cosine_similarity(&emb1, &emb2);
assert!((sim - 1.0).abs() < 1e-6);
let emb3 = Array1::from_vec(vec![0.0, 1.0, 0.0]);
let sim2 = encoder.cosine_similarity(&emb1, &emb3);
assert!(sim2.abs() < 1e-6);
}
}
|